wiki: 2026-11-14 audit corrections — withdraw the 'sign not magnitude' / '12 h of lead time' / '18:03 run' claims; verified escalation mechanism (activity-aware magnitude gate) + 29-cycle zero-item consolidation streak

This commit is contained in:
Tony0410
2026-11-15 01:28:53 +08:00
parent 2d439145bf
commit 82876abdfb
5 changed files with 134 additions and 31 deletions
+2 -2
View File
@@ -14,7 +14,7 @@ sources: [live-cron-ledger, live-deliveries-db, live-incident-ledger, live-sessi
> Operational knowledge for Anthony's agents, infrastructure,
> model routing and automation. Last updated: 2026-11-15.
>
> **2026-11-14 (nightly audit; live probes 2026-11-14 23:59 - 2026-11-15 00:4x AWST) — a third TTS stack landed on CT200, and the memory pipeline went quiet.** **New page [[systems/phonon-tts]]**: `ptts-openai-server` (`:8881`, OpenAI-compatible `/v1/audio/speech`, model **`gradium/phonon-40m-en`**, container created 2026-11-14 10:31:56 AWST), `phonon-demo` (`:8882`, WASM demo UI on `nginx:alpine`, 11:00:11), `ptts-ws-server` (`:8883`, WebSocket, locally-built `ptts-ws-server:0.7.1`, 11:43:59) — verified live from CT460: `/health` 200, `/v1/models`, and a `POST /v1/audio/speech` returning a valid MP3; `/opt/ptts-voices` holds **7** voices; both ptts containers were restarted together at **17:54 AWST**. ⚠ **Provenance gap**: the retained transcripts cover only 08:35–09:03 AWST, so the 10:31 / 11:00 / 11:43 / 17:54 changes are **not attributable to any session in this profile** (working hypothesis: another estate agent — unproven). **New open incident [[incidents/2026-11-14-mnemosyne-capture-stall]]**: Mnemosyne has captured **nothing since 2026-11-14 09:32:59 AWST** — `working_memory` and `episodic_memory` stop in the same minute, last 8 working writes all 01:01–01:32 UTC — and the health monitor `ea79a2107e36` **escalated only at 2026-11-15 00:03:35** (exit 1, incident `ea79a2_601037516abd`, `failure_type: auth` — **mislabelled**, state **`alerted`**, unclosed, alert text printing the raw **UTC** value so it reads as a 23 h gap when the true AWST gap was 14 h 31 m). ⚠ **~12 h of lead time produced no alarm**: the 12:00:52 AWST run saw the same hole and logged it **ℹ️ informational** (`capture delay -589.5 min` — the negative sign hides the growth). ~15 h of desktop (last message 19:22 AWST) / Telegram / cron context is uncaptured, and the **23:30 sleep cycle consolidated 0 items**. Ruled out: the known stringified-config killer (`config.yaml` mtime **2026-11-13 18:03**; full YAML walk found no `[`/`{` scalars). Root cause **not established**. **Fleet otherwise clean — until the monitor's own 00:03 alert: 2026-11-14 settled 167 rows — 166 `completed` / 0 `failed` / 1 `running`** (the running row is this audit), deliveries **4 / 4 `delivered`**, no other incident row moved, so `cron_incidents` went **0 → 1 unclosed** (the monitor's row, `alerted_at` 00:03:44). 2026-11-15 so far: **8 rows — 7 `completed` / 1 `failed`** (the monitor), 1 delivery `delivered`. Fleet **29 (26 / 3)** — a **ninth** window unchanged, but **non-clean `last_status` 0 → 1** (`ea79a2107e36`, `error`, `failure_streak` 1), the first red row since the Uber Eats job cleared on 2026-11-13. [[infrastructure/docker-services]] re-verified: CT200 **64 running non-seeder / 72 total** excluding the 1 transient seeder (was 61 / 68 on 11-07) — the delta is **exactly** the three TTS containers, **no removals**. The **2026-11-08 hook cascade** gained **no** new occurrence in the 11-14 body (the 00:05:40 timeout was already recorded by the 11-13 audit; `errors.log` 11-14 = 1 timeout + 10 suppression lines, otherwise only the benign `Browser Use viewing unavailable` plugin warning at ~2/hour).
> **2026-11-14 (nightly audit; live probes 2026-11-14 23:59 - 2026-11-15 00:4x AWST) — a third TTS stack landed on CT200, and the memory pipeline went quiet.** **New page [[systems/phonon-tts]]**: `ptts-openai-server` (`:8881`, OpenAI-compatible `/v1/audio/speech`, model **`gradium/phonon-40m-en`**, container created 2026-11-14 10:31:56 AWST), `phonon-demo` (`:8882`, WASM demo UI on `nginx:alpine`, 11:00:11), `ptts-ws-server` (`:8883`, WebSocket, locally-built `ptts-ws-server:0.7.1`, 11:43:59) — verified live from CT460: `/health` 200, `/v1/models`, and a `POST /v1/audio/speech` returning a valid MP3; `/opt/ptts-voices` holds **7** voices; both ptts containers were restarted together at **17:54 AWST**. ⚠ **Provenance gap**: the retained transcripts cover only 08:35–09:03 AWST, so the 10:31 / 11:00 / 11:43 / 17:54 changes are **not attributable to any session in this profile** (working hypothesis: another estate agent — unproven). **New open incident [[incidents/2026-11-14-mnemosyne-capture-stall]]**: Mnemosyne has captured **nothing since 2026-11-14 09:32:59 AWST** — `working_memory` and `episodic_memory` stop in the same minute, last 8 working writes all 01:01–01:32 UTC — and the health monitor `ea79a2107e36` **escalated only at 2026-11-15 00:03:35** (exit 1, incident `ea79a2_601037516abd`, `failure_type: auth` — the scheduler's classifier, not a credential fault, state **`alerted`**, unclosed, alert text printing the raw **UTC** value so it reads as a 23 h gap when the true AWST gap was 14 h 31 m). ⚠ **Correction, re-derived 2026-11-15 01:2x: the 12:00:52 run saw nothing** — it measured `capture delay 29.4 min` (healthy, exit 0, `suppressed`), and the `-589.5 min` ℹ️ line belongs to the **00:03:35 run's own output block** (monitor log: header 957 = 12:01:52 run, header 1030 = 00:03:35 run, `-589.5` at 1046, ❌ at 1047). The earlier "~12 h of lead time" claim is **withdrawn**, as is the "exited 0 at 12:00 and 18:03" sentence (schedule `0 */12`; there was no 18:03 slot). Detection is **activity-aware** — `mnemosyne_freshness_gate.py` fires on `delay < −120 min` (magnitude-based, no sign-only test) and can only fire once a conversational message postdates the last capture; the conversation was idle 09:03→19:22, so midnight was the first slot able to see the miss. **What actually went unread: 29 consecutive 0-item consolidation cycles** from **10:04:42 AWST**, printed as `(0 items)` every run yet tested only for age (`> 36 h`). ~15 h of desktop (last message 19:22 AWST) / Telegram / cron context is uncaptured, and the **23:30 sleep cycle consolidated 0 items**. Ruled out: the known stringified-config killer (`config.yaml` mtime **2026-11-13 18:03**; full YAML walk found no `[`/`{` scalars). Root cause **not established**. **Fleet otherwise clean — until the monitor's own 00:03 alert: 2026-11-14 settled 167 rows — 166 `completed` / 0 `failed` / 1 `running`** (the running row is this audit), deliveries **4 / 4 `delivered`**, no other incident row moved, so `cron_incidents` went **0 → 1 unclosed** (the monitor's row, `alerted_at` 00:03:44). 2026-11-15 so far: **8 rows — 7 `completed` / 1 `failed`** (the monitor), 1 delivery `delivered`. Fleet **29 (26 / 3)** — a **ninth** window unchanged, but **non-clean `last_status` 0 → 1** (`ea79a2107e36`, `error`, `failure_streak` 1), the first red row since the Uber Eats job cleared on 2026-11-13. [[infrastructure/docker-services]] re-verified: CT200 **64 running non-seeder / 72 total** excluding the 1 transient seeder (was 61 / 68 on 11-07) — the delta is **exactly** the three TTS containers, **no removals**. The **2026-11-08 hook cascade** gained **no** new occurrence in the 11-14 body (the 00:05:40 timeout was already recorded by the 11-13 audit; `errors.log` 11-14 = 1 timeout + 10 suppression lines, otherwise only the benign `Browser Use viewing unavailable` plugin warning at ~2/hour).
> **2026-11-13 (nightly audit; live probes 2026-11-13 23:59 - 2026-11-14 00:2x AWST) — the fleet went fully clean, then the hook came back.** **2026-11-13: 174 ledger rows — 173 `completed` / 0 `failed` / 1 `running`** (this audit, started 23:59:41) — the **third consecutive zero-`failed` day** (11-11 172/172, 11-12 176/176); deliveries **8 / 8 `delivered`**, 0 `for_failure`; `cron_incidents` **0 unclosed**. **Every job now reads `last_status: ok` with `failure_streak` 0 — the fleet has no non-clean row at all**, because **`066bb5191280` (Uber Eats Monthly Spend Report) was fixed end-to-end on 2026-11-13**, clearing a `delivery_failed` that had stood since 2026-11-01. Two stacked causes: the job's **runtime prompt had grown to 88 k chars** (it carried two heavy browser skills) so the 2026-11-01 run exhausted its turn budget mid-format and shipped a truncated digest, and the `delivery_failed` label was a **stale `TELEGRAM_HOME_CHANNEL_THREAD_ID=34795` env var** that only clears on a gateway restart (**2026-11-13 13:14**). Fix: deterministic printer **`~/.hermes/cron/scripts/uber-eats-digest.py`**, rewritten prompt with a hard output contract, skills cleared (**88 k -> 1.8 k**); verified on the 18:51 -> 18:58 slot (**154 unique / 153 completed** orders, Oct **9 orders / $301.04**), lesson saved to `cron-job-maintenance`. Fleet **29 (26 / 3)** — an **eighth** window unchanged. ⚠ **But the window's fault is the `pre_tool_call` hook, and it returned:** `_pre_tool_call timed out after 120s` at **2026-11-14 00:05:40** plus **10** suppression lines through **00:06:38**, refusing this audit's own tool calls for ~1 minute (first-hand). `errors.log` by day: 11-12 **0/0**, 11-13 **0/0**, 11-14 **1/10** — the first such timeout in the retained log since the 2026-11-08 cascade (**`errors.log` now starts 2026-11-09 13:18:39**, so the 11-08 lines have rotated out and the page's 58-count stands on the 2026-11-09 read). **Load is the trigger and it has ~doubled: `load average 12.08 / 13.02 / 12.33` on 2 vCPU** (vs **5.89** at diagnosis), with **`state.db` now 2.04 GB** (**2,977 sessions / 207,577 messages**, was 1.89 GB / 2,901 / 197,682) and the root filesystem **88 % used**. The 30 -> 120 s fix is **proven insufficient**; [[incidents/2026-11-08-pre-tool-call-hook-timeout-cascade]] is **reopened, `resolved` -> `active`**, and `state.db` checkpoint/retention — not a bigger timeout — is now the estate's most urgent open item. Freshness (03:00 stocktake): **0 stale / 131 fresh / 3 no-date of 134**.
>
> **2026-11-11 → 2026-11-12 (nightly audit; live probes 2026-11-12 23:59 – 2026-11-13 00:4x AWST) — a clean day, and an upgrade that could not be installed.** The 429 row that was the estate's only open incident **closed `resolved` at 2026-11-12T00:25:36** (the next slot of the same job ran clean, carrying both windows); *cron_incidents* now reads **0 unclosed of 137**. **2026-11-12: 176 ledger rows — 175 `completed` / 0 `failed` / 1 `running`** (this audit, started 23:59:54), deliveries **9 / 9 `delivered`**; **11-11: 172 / 172 `completed` / 0 `failed`**, deliveries 7 / 7. Fleet **29 (26 / 3)**, a seventh window unchanged. The window's one new open item is not a failure: the attempt to take **`hermes-hq-push`** past its installed **0.9.42** was **refused by the pre-install plugin security scan** — *"Blocked (dangerous verdict, 1 critical of 39 findings (echo_pipe_exec)). --force does not override a dangerous verdict."* The critical is the plugin's **own** `skills/dashboard-tile/SKILL.md` one-liner; **no per-verdict override exists** (`plugins.scan_on_install: false` would blind the scan for *every* future install, and `plugin_guard.py` hardcodes `trust_level="community"`). The fix is **PR #2** on `mrcharlesiv/hermes-hq-plugins` (author `Tony0410`, `fix/tile-skill-heredoc`, **OPEN / MERGEABLE**, `mergedAt: null`, +7/−2 in one file), pushed from a fork because `gh` is **read-only** upstream. Installed 0.9.42 keeps delivering. Also: the alleged *"stored hold-and-ask instruction"* for scan warnings **does not exist on disk** (184,716-file sweep, zero matches in editable text; the sentence lives only in `state.db` history) — nothing was removed or reverted, and the hunt is now the skill `stored-instruction-search-and-removal` (**skills 299 → 300**). `config.yaml` is **byte-identical to its 2026-11-12 18:14:52 backup**; the only drift since 11-09 is the **model catalogue block** (6 out, 9 in). **New page:** [[incidents/2026-11-12-hq-push-upgrade-blocked-by-plugin-scan]].
@@ -80,7 +80,7 @@ sources: [live-cron-ledger, live-deliveries-db, live-incident-ledger, live-sessi
- [[systems/skills-index]] — Installed skills inventory (**300 skills** (**67 categories**) as of the **2026-11-13** audit; regenerated daily by the 03:00 job). The generator **excludes `.archive/`** (50 archived skills) — the earlier 346/347 counts included it. ⚠ **2026-11-07 07:43 — four skills moved into `.archive/` in one pass** (`cron-model-repair`, `fleet-wide-failure-triage`, `hermes-instance-update`, `mail-lookup`), taking the count **302 → 299** and the parked archive **46 → 50**; the `Cron Model Repair` category went with its only member. **No active `SKILL.md` references any of the four**, so if this was a consolidation the *"absorbed into X"* pointer is missing. One skill appeared in the same window (`microsoft-office`, Productivity 25 → 26). Freshness at the **2026-11-08** run: **0 stale / 126 fresh / 3 no-date of 129 pages** (`freshness-20261108.txt`). Note: `SKILL.md` files on disk still exceed the index by one (300 vs 299) — the generator de-dups one nested duplicate, so **299 is the correct unique-skill figure**; see [[systems/scheduled-tasks]]. ✅ **2026-11-05 — the generator divergence that made this page flip-flop is FIXED at the root:** two *different* scripts were writing this page (the repo script → 302/68; the stale `~/.hermes/scripts/wiki-maintain.sh` copy → 349/26, because it used `Path.rglob()` and so counted the parked archive), and the nightly audit was pointed at the stale one. The stale copy is now a **delegating wrapper** with no generator logic of its own → [[incidents/2026-11-05-skills-index-generator-divergence]]. ⚠ The **2026-10-22** slot ran **97 m 58 s** (vs 16 m 01 s on 10-21), finished `completed` with delivery `suppressed` and **committed nothing**; the 10-23 slot closed that hole.
## Incidents
- [[incidents/2026-11-14-mnemosyne-capture-stall]] - ⚠ **OPEN — Mnemosyne has captured nothing since 2026-11-14 09:32:59 AWST (~15 h at this read), and the health monitor took 14 h 31 m to say so.** `working_memory` and `episodic_memory` stop in the same minute (last 8 working writes all 01:01–01:32 UTC); the `pre_llm_call` capture hook logs `Mnemosyne capture ok` on every desktop turn while writing **0** rows, so the hook's own success is not evidence of capture. The monitor `ea79a2107e36` escalated at **2026-11-15 00:03:35** (exit 1, incident `ea79a2_601037516abd`, `failure_type: auth` — **mislabelled**, state **`alerted`**, unclosed), and its alert text prints the raw **UTC** instant, so it reads as a 23 h gap when the true AWST gap was 14 h 31 m. ⚠ **The 12:00:52 AWST run saw the identical hole and logged it ℹ️ informational** (`capture delay -589.5 min`), i.e. **~12 h of lead time produced no alarm**. Consequences: ~15 h of desktop / Telegram / cron context uncaptured, and the **23:30 sleep cycle consolidated 0 items**. Ruled out: the known stringified-config killer (`config.yaml` mtime **2026-11-13 18:03**; full YAML walk found no `[`/`{` scalars). Root cause **not established** → [[systems/mnemosyne]], [[systems/phonon-tts]].
- [[incidents/2026-11-14-mnemosyne-capture-stall]] - ⚠ **OPEN — Mnemosyne has captured nothing since 2026-11-14 09:32:59 AWST (~15 h at this read), and the health monitor took 14 h 31 m to say so.** `working_memory` and `episodic_memory` stop in the same minute (last 8 working writes all 01:01–01:32 UTC); the `pre_llm_call` capture hook logs `Mnemosyne capture ok` on every desktop turn while writing **0** rows, so the hook's own success is not evidence of capture. The monitor `ea79a2107e36` escalated at **2026-11-15 00:03:35** (exit 1, incident `ea79a2_601037516abd`, `failure_type: auth` — the scheduler's classifier, not a credential fault; this job's twelve earlier rows read `agent` / `timeout` / `unknown` — state **`alerted`**, unclosed), and its alert text prints the raw **UTC** instant, so it reads as a 23 h gap when the true AWST gap was 14 h 31 m. ⚠ **Correction, re-derived 2026-11-15 01:2x: the 12:00:52 AWST run saw nothing** (`capture delay 29.4 min`, exit 0, `suppressed`) — the `-589.5 min` ℹ️ line belongs to the **00:03:35** run itself (monitor log lines 957 / 1030 / 1046). Detection is **activity-aware** (`mnemosyne_freshness_gate.py` fires on `delay < −120 min`, i.e. on **magnitude**) and the first message to postdate the stall arrived at **19:22 AWST**, after that slot — so the ~12 h of quiet was the `0 */12` **cadence**, not a threshold bug. **What went unread was volume: 29 consecutive 0-item consolidation cycles from 10:04:42 AWST**, visible as `(0 items)` in every run but tested only for age (36 h). Consequences: ~15 h of desktop / Telegram / cron context uncaptured, and the **23:30 sleep cycle consolidated 0 items**. Ruled out: the known stringified-config killer (`config.yaml` mtime **2026-11-13 18:03**; full YAML walk found no `[`/`{` scalars). Root cause **not established** → [[systems/mnemosyne]], [[systems/phonon-tts]].
- [[incidents/2026-11-12-hq-push-upgrade-blocked-by-plugin-scan]] - **OPEN — the `hermes-hq-push` upgrade past v0.9.42 was refused by the pre-install plugin security scan.** `hermes plugins install mrcharlesiv/hermes-hq-plugins/hermes-hq-push --enable --force` returned **"Blocked (dangerous verdict, 1 critical of 39 findings (echo_pipe_exec)). --force does not override a dangerous verdict."** The sole critical is the plugin's **own** `skills/dashboard-tile/SKILL.md:41` shell one-liner (`echo '<json>' | python3 …` = *obfuscation*); the other 38 findings are low/medium and alone would have installed as `caution`. **`--force` lifts `caution` only**; a `dangerous` verdict is an unconditional block (`tools/plugin_guard.py:406-418`), and **no per-verdict override exists** — the only key is `plugins.scan_on_install` (default `true`, currently unset), which disables scanning for *every* future install; the `trust_level`/allowlist machinery lives in `tools/skills_guard.py` (the skills path) and `plugin_guard.py` hardcodes `trust_level="community"`. **Installed state unchanged:** v0.9.42, `enabled`, still delivering. **Fix = PR #2** on `mrcharlesiv/hermes-hq-plugins` (author `Tony0410`, `fix/tile-skill-heredoc` → `main`, **OPEN / MERGEABLE**, `mergedAt: null`, +7/−2 in one file, created 2026-11-12T10:00:01Z) — pushed from a fork because this host's `gh` account has **read-only** access upstream; local clone `/home/hermes/workspace/hq-plugins-fix` (`90c9d28`).
- [[incidents/2026-11-10-wiki-audit-429-all-targets-skipped]] - **RESOLVED 2026-11-12T00:25:36** (self-cleared on the next slot's clean completion; *cron_incidents* now 0 unclosed of 137) **— the auditor failed, so 2026-11-10 was never audited.** `cbfe434a3b5e` (Wiki Audit — Nightly) ran **2026-11-10T23:59:54.486572 -> 2026-11-11T00:12:29.646987 = 12 m 35 s** and died on `RuntimeError: HTTP 429: Service temporarily unavailable: all targets were skipped by pre-dispatch filters (credits 100% conn:3021bab6) reset after 6h 24m 31s` (delivery `suppressed`) — **6 targets evaluated, `attempted 0`**, tenant `tn-5ud77abtl1` quota exceeded. Incident row `cbfe43_ed01845310d8` (`rate_limit`) was raised at 00:12:28 and is **still open with `alerted_at: NULL` — the estate's only open incident row**. Fleet non-clean `last_status` **1 -> 2**; pool reset ≈ **06:36:57**. Third symptom of the same router behaviour (cf. [[incidents/2026-10-21-omniroute-pre-dispatch-filter-skip]]).
- [[incidents/2026-11-08-pre-tool-call-hook-timeout-cascade]] - ⚠ **REOPENED 2026-11-14 — the hook was the fault, and the 30 → 120 s fix did not hold at current load.** The `conduit_push` plugin is the **only** `pre_tool_call` registrant on the host (`__init__.py:119`) and its callback overran Hermes's **30 s** hook budget **58 times** between **2026-11-07 14:14:03** and **2026-11-08 08:21:03** (`errors.log`: **12** lines on 11-07, **46** on 11-08) plus **89** *"skipped after previous timeout or while still running"* suppression lines — every tool call degraded to slow-then-suppressed. **Fixed 2026-11-08 13:35 AWST: `plugins.hook_callback_timeout` 30 -> 120** in `~/.hermes/config.yaml`; the last timeout line remains 08:21:03 (re-verified at audit time). ⚠ **The plugin's caching was measured, not assumed, and was innocent** — `session_titles.py:21 CACHE_SECONDS = 30.0`, timed in-process at **3.1 s cold / 0.00 s warm**, so the *"re-opens a 1.87 GB `state.db` per call, ~9 s each"* theory was **wrong**; the trigger is **load** on the 2-core CT460 (`load average 5.89`, `state.db` checkpointing a **40 MB WAL**). **Cost:** *Mnemosyne Health Monitor* **8 h 27 m** and **5 h 59 m 51 s**; *FreshRSS Midday* **2 h 09 m 53 s**; stocktake failed a third day. **Next lever is `state.db` checkpoint/retention, not more caching.** ⚠ **Recurred 2026-11-14 00:05:40** — **1 timeout + 10 skips**, overrunning the **120 s** budget at `load average 12.08` on 2 vCPU (**~2×** the 5.89 at diagnosis), with `state.db` grown to **2.04 GB**. Status **`resolved` → `active`**; the lever is now urgent. → [[systems/conduit]], [[systems/hermes-agent]]